---
title: "How AI agents actually make decisions (and where human sign-off matters)"
description: "AI agents read, decide, draft and act. At Automancer a human signs off before money moves or anything client-facing goes out."
url: "https://automancer.uk/field-notes/how-ai-agents-actually-make-decisions/"
canonical: "https://automancer.uk/field-notes/how-ai-agents-actually-make-decisions/"
type: "BlogPosting"
author: "Waseem Ilyas"
datePublished: "2026-09-07T00:00:00.000Z"
---
"AI agent" gets used for everything from a chatbot with a new coat of paint
to something that actually runs part of a business. Here is the plain
version, with no hand-waving: what an agent does, what "governed" means in
practice, and exactly where we still put a human in front of the decision.

## What is an AI agent, actually?

An AI agent is software that reads, decides, drafts and acts, on its own,
inside limits someone else set. That is the whole definition, and it is
different from the two things people usually mean by "AI" in a small
business context. A chatbot answers questions; it does not do anything. A
rules-only script does something, but only the one thing it was told to do,
the same way, every time. An agent sits between them: it can look at a
messy, real situation, decide what the right next step is, and carry it out,
the same way a competent junior member of staff would.

That is the useful part and the scary part in the same sentence. Software
that can decide things is only as trustworthy as the process around it. Which
is the actual subject of this post.

## What "governed" means in practice

Governed does not mean "approved for every step by a human before it moves."
That would just be automation with extra paperwork, and most of the value
disappears with the speed. Governed means the work moves through a deploy
loop, not a permission queue:

- **Scoped and tested in staging** before it touches anything real.
- **Reconciled against real data**, so the agent's version of the world is
  checked against what is actually true, not just what it assumed.
- **Monitored in production**, so a problem gets caught while it is small.
- **Verified after deployment**, so "it ran" and "it ran correctly" are
  checked as two separate questions.

Notice what is missing from that list: a person clicking "approve" on every
individual action. That is deliberate. The safety comes from the loop being
tight and honest, monitored and reconciled and reversible, not from a queue
of sign-offs that a busy human will eventually start rubber-stamping anyway.

## Where the line actually sits

We run our own business this way, so the rule is not theoretical: agents
make routine, scoped and reversible decisions on their own. A human signs off
before **money** moves or anything **client-facing** goes out. Drafting an
invoice, monitoring a deployment, flagging a stalled task, deciding which of
two systems has the correct number today, agents do all of that without
asking first, because getting it wrong is cheap to catch and cheap to
reverse. Sending that invoice, publishing something under our name, or
committing to a price or a promise a client will hold us to, a person signs
off first, because getting those wrong is expensive and sometimes impossible
to take back.

## Why draw the line there, specifically

The honest reason is that not every action carries the same cost when it
goes wrong. A draft sitting in a queue for review costs nothing extra to
produce and nothing to discard if it is wrong. Money moving, or something
going out under a business's name, is a different category: it touches
someone else's bank balance or someone else's inbox, and undoing it is
slower and more awkward than getting it right the first time.

The alternative, a sign-off gate on every irreversible-feeling action,
sounds safer and is not. It just relocates the risk from "the agent got
something wrong" to "the human stopped actually reading what they were
approving," because nobody sustains careful attention across a hundred
low-stakes approvals a day. Two gates that matter, watched properly, beat
twenty gates that get waved through on autopilot. That is the actual trade,
and it is why we picked money and client-facing action as the two places a
human always looks before anything moves.

## What this means if you are evaluating an AI vendor

Ask them the same question we just answered about ourselves: which
decisions does the software make on its own, and which ones always stop for
a person first. A vendor who cannot answer that quickly, or who says
"everything gets reviewed" as if that were a feature, has not actually
thought about where the risk lives. The right answer sounds specific: these
kinds of actions run themselves inside a tested, monitored loop; these other
kinds always wait for a human. If they cannot draw that line for their own
product, they have not drawn it for yours either.

This is what governed AI actually looks like: not a human in every step, but
a human at the two steps that matter, and a tight, monitored, reconciled
loop for everything in between. If you want to see what that looks like on a
real system rather than a description of one, [how we work](/about/) covers
the rest of it, or an [Automation Opportunity Audit](/services/) is where we
map the same question onto your business.